paper

official tensorflow version

This model is extreamly useful for real-world RE usage. I originally reimplemented for a competition (Chinese IE). I will add CoNLL04 dataset and BERT model.

• python 3.6
• pytorch 1.20

Chinese Information Extraction Competition link

Unzip *.json into ./raw_data/chinese/

We use the data processed by official version.

already in ./raw_data/CoNLL04/

python main.py --mode preprocessing --exp_name chinese_selection_re
python main.py --mode train --exp_name chinese_selection_re 
python main.py --mode evaluation --exp_name chinese_selection_re

If you want to try other experiments:

set exp_name as conll_selection_re or conll_bert_re

Training speed: 10min/epoch

Test set:

We use the strictest setting: a triplet is correct only if the relation and all the tokens of head and tail are correct.

The model was originally used for Chinese IE, thus, it's a bit different from the official paper:

They use pretrained char-word embedding while we use word embedding initialized randomly; they use 3-layer LSTM while we use 1-layer LSTM.

offline bert model fix bugs of bert tokenizer update and add all requirements

• Tune the hyperparameters of BERT+MUL for Chinese
• Add full fine-tune bert